def compress(self, inputs):
"""Compress inputs and store their binary representations into strings.
Args:
inputs: `Tensor` with values to be compressed.
Returns:
String `Tensor` vector containing the compressed representation of each
batch element of `inputs`.
"""
with ops.name_scope(self._name_scope()):
inputs = ops.convert_to_tensor(inputs)
if not self.built:
# Check input assumptions set before layer building, e.g. input rank.
input_spec.assert_input_compatibility(self.input_spec, inputs, self.name)
if self.dtype is None:
self._dtype = inputs.dtype.base_dtype.name
self.build(inputs.shape)
# Check input assumptions set after layer building, e.g. input shape.
if not context.executing_eagerly():
input_spec.assert_input_compatibility(self.input_spec, inputs, self.name)
ndim = self.input_spec.ndim
channel_axis = self._channel_axis(ndim)
# Tuple of slices for expanding dimensions of tensors below.
slices = ndim * [None] + [slice(None)]
slices[channel_axis] = slice(None)
slices = tuple(slices)
# Expand dimensions of CDF to input dimensions, keeping the channels along
# the right dimension.
cdf = self._quantized_cdf[slices[1:]]
num_levels = array_ops.shape(cdf)[-1] - 1
# Bring inputs to the right range by centering the range on the medians.
half = constant_op.constant(.5, dtype=self.dtype)
medians = array_ops.squeeze(self._medians, [1, 2])
offsets = (math_ops.cast(num_levels // 2, self.dtype) + half) - medians
# Expand offsets to input dimensions and add to inputs.
values = inputs + offsets[slices[:-1]]
# Clip to range and cast to integers. Because we have added .5 above, and
# all values are positive, the cast effectively implements rounding.
values = math_ops.maximum(values, half)
values = math_ops.minimum(
values, math_ops.cast(num_levels, self.dtype) - half)
values = math_ops.cast(values, dtypes.int16)
def loop_body(tensor):
return coder_ops.range_encode(
tensor, cdf, precision=self.range_coder_precision)
strings = functional_ops.map_fn(
loop_body, values, dtype=dtypes.string, back_prop=False)
if not context.executing_eagerly():
strings.set_shape(inputs.shape[:1])
return strings
def compress(self, inputs, input_stddev):
"""Compress inputs and store their binary representations into strings.
Args:
inputs: `Tensor` with values to be compressed.
Returns:
String `Tensor` vector containing the compressed representation of each
batch element of `inputs`.
"""
with ops.name_scope(self._name_scope()):
inputs = ops.convert_to_tensor(inputs)
if not self.built:
# Check input assumptions set before layer building, e.g. input rank.
input_spec.assert_input_compatibility(self.input_spec, inputs,
self.name)
if self.dtype is None:
self._dtype = inputs.dtype.base_dtype.name
self.build(inputs.shape)
input_stddev = ops.convert_to_tensor(input_stddev)
inputs = array_ops.expand_dims(inputs, axis=4)
input_stddev = array_ops.expand_dims(input_stddev, axis=4)
self.build_gauss(input_stddev)
return tf.zeros(shape=inputs.shape[:1],dtype=tf.string)
# Check input assumptions set after layer building, e.g. input shape.
# Expand dimensions of CDF to input dimensions, keeping the channels along
# the right dimension.
cdf = self._quantized_cdf
num_levels = array_ops.shape(cdf)[-1] - 1
half_num_levels = math_ops.cast(num_levels // 2,self.dtype)
# Bring inputs to the right range by centering the range on the medians.
half = constant_op.constant(.5, dtype=self.dtype)
offsets = - self._medians + ( half_num_levels + half)
# Expand offsets to input dimensions and add to inputs.
values = inputs + offsets
# Clip to range and cast to integers. Because we have added .5 above, and
# all values are positive, the cast effectively implements rounding.
values = math_ops.maximum(values, half)
values = math_ops.minimum(
values, math_ops.cast(num_levels, self.dtype) - half)
values = math_ops.cast(values, dtypes.int16)
values = array_ops.squeeze(values, [-1])
def loop_body(tensor):
return coder_ops.range_encode(
tensor, cdf, precision=self.range_coder_precision)
strings = functional_ops.map_fn(
loop_body, values, dtype=dtypes.string, back_prop=False)
if not context.executing_eagerly():
strings.set_shape(inputs.shape[:1])
return strings
def _maybe_build(self, inputs):
# Check input assumptions set before layer building, e.g. input rank.
if not self.built:
input_spec.assert_input_compatibility(self.input_spec, inputs,
self.name)
input_list = nest.flatten(inputs)
input_shapes = None
if all(hasattr(x, 'shape') for x in input_list):
input_shapes = nest.map_structure(lambda x: x.shape, inputs)
# Only call `build` if the user has manually overridden the build method.
if not hasattr(self.build, '_is_default'):
# Any setup work performed only once should happen in an `init_scope`
# to avoid creating symbolic Tensors that will later pollute any eager
# operations.
with tf_utils.maybe_init_scope(self):
self.build(input_shapes)
# We must set self.built since user defined build functions are not
# constrained to set self.built.
self.built = True
def compress(self, inputs):
"""Compress inputs and store their binary representations into strings.
Arguments:
inputs: `Tensor` with values to be compressed.
Returns:
compressed: String `Tensor` vector containing the compressed
representation of each batch element of `inputs`.
Raises:
ValueError: if `inputs` has an integral or inconsistent `DType`, or
inconsistent number of channels.
"""
with tf.name_scope(self._name_scope()):
inputs = tf.convert_to_tensor(inputs, dtype=self.dtype)
if not self.built:
# Check input assumptions set before layer building, e.g. input rank.
input_spec.assert_input_compatibility(self.input_spec, inputs,
self.name)
if self.dtype is None:
self._dtype = inputs.dtype.base_dtype.name
self.build(inputs.shape)
# Check input assumptions set after layer building, e.g. input shape.
if not tf.executing_eagerly():
input_spec.assert_input_compatibility(self.input_spec, inputs,
self.name)
if inputs.dtype.is_integer:
raise ValueError("{} can't take integer inputs.".format(
type(self).__name__))
symbols = self._quantize(inputs, "symbols")
assert symbols.dtype == tf.int32
ndim = self.input_spec.ndim
indexes = self._prepare_indexes(shape=tf.shape(symbols)[1:])
broadcast_indexes = (indexes.shape.ndims != ndim)
if broadcast_indexes:
# We can't currently broadcast over anything else but the batch axis.
assert indexes.shape.ndims == ndim - 1
args = (symbols, )
else:
args = (symbols, indexes)
def loop_body(args):
string = range_coding_ops.unbounded_index_range_encode(
args[0],
indexes if broadcast_indexes else args[1],
self._quantized_cdf,
self._cdf_length,
self._offset,
precision=self.range_coder_precision,
overflow_width=4,
debug_level=0)
return string
strings = tf.map_fn(loop_body,
args,
dtype=tf.string,
back_prop=False,
name="compress")
if not tf.executing_eagerly():
strings.set_shape(inputs.shape[:1])
return strings
def compress(self, inputs):
"""Compress inputs and store their binary representations into strings.
Arguments:
inputs: `Tensor` with values to be compressed.
Returns:
compressed: String `Tensor` vector containing the compressed
representation of each batch element of `inputs`.
Raises:
ValueError: if `inputs` has an integral or inconsistent `DType`, or
inconsistent number of channels.
"""
with tf.name_scope(self._name_scope()):
inputs = tf.convert_to_tensor(inputs, dtype=self.dtype)
if not self.built:
# Check input assumptions set before layer building, e.g. input rank.
input_spec.assert_input_compatibility(
self.input_spec, inputs, self.name)
if self.dtype is None:
self._dtype = inputs.dtype.base_dtype.name
self.build(inputs.shape)
# Check input assumptions set after layer building, e.g. input shape.
if not tf.executing_eagerly():
input_spec.assert_input_compatibility(
self.input_spec, inputs, self.name)
if inputs.dtype.is_integer:
raise ValueError(
"{} can't take integer inputs.".format(type(self).__name__))
symbols = self._quantize(inputs, "symbols")
assert symbols.dtype == tf.int32
ndim = self.input_spec.ndim
indexes = self._prepare_indexes(shape=tf.shape(symbols)[1:])
broadcast_indexes = (indexes.shape.ndims != ndim)
if broadcast_indexes:
# We can't currently broadcast over anything else but the batch axis.
assert indexes.shape.ndims == ndim - 1
args = (symbols,)
else:
args = (symbols, indexes)
def loop_body(args):
string = range_coding_ops.unbounded_index_range_encode(
args[0], indexes if broadcast_indexes else args[1],
self._quantized_cdf, self._cdf_length, self._offset,
precision=self.range_coder_precision, overflow_width=4,
debug_level=0)
return string
strings = tf.map_fn(
loop_body, args, dtype=tf.string,
back_prop=False, name="compress")
if not tf.executing_eagerly():
strings.set_shape(inputs.shape[:1])
return strings
def inference(self, inputs, *args, **kwargs):
call_context = base_layer_utils.call_context()
input_list = nest.flatten(inputs)
# We will attempt to build a TF graph if & only if all inputs are symbolic.
# This is always the case in graph mode. It can also be the case in eager
# mode when all inputs can be traced back to `keras.Input()` (when building
# models using the functional API).
build_graph = tf_utils.are_all_symbolic_tensors(input_list)
# Accept NumPy and scalar inputs by converting to Tensors.
if any(isinstance(x, (np.ndarray, float, int)) for x in input_list):
def _convert_non_tensor(x):
# Don't call `ops.convert_to_tensor` on all `inputs` because
# `SparseTensors` can't be converted to `Tensor`.
if isinstance(x, (np.ndarray, float, int)):
return ops.convert_to_tensor(x)
return x
inputs = nest.map_structure(_convert_non_tensor, inputs)
input_list = nest.flatten(inputs)
# Handle `mask` propagation from previous layer to current layer. Masks can
# be propagated explicitly via the `mask` argument, or implicitly via
# setting the `_keras_mask` attribute on the inputs to a Layer. Masks passed
# explicitly take priority.
mask_arg_passed_by_framework = False
input_masks = self._collect_input_masks(inputs, args, kwargs)
if (self._expects_mask_arg and input_masks is not None and
not self._call_arg_was_passed('mask', args, kwargs)):
mask_arg_passed_by_framework = True
kwargs['mask'] = input_masks
# If `training` argument was not explicitly passed, propagate `training`
# value from this layer's calling layer.
training_arg_passed_by_framework = False
# Priority 1: `training` was explicitly passed.
if self._call_arg_was_passed('training', args, kwargs):
training_value = self._get_call_arg_value('training', args, kwargs)
if not self._expects_training_arg:
kwargs.pop('training')
else:
training_value = None
# Priority 2: `training` was passed to a parent layer.
if call_context.training is not None:
training_value = call_context.training
# Priority 3a: `learning_phase()` has been set.
elif backend.global_learning_phase_is_set():
training_value = backend.learning_phase()
# Priority 3b: Pass the `learning_phase()` if in the Keras FuncGraph.
elif build_graph:
with backend.get_graph().as_default():
if base_layer_utils.is_in_keras_graph():
training_value = backend.learning_phase()
if self._expects_training_arg and training_value is not None:
# Force the training_value to be bool type which matches to the contract
# for layer/model call args.
if tensor_util.is_tensor(training_value):
training_value = math_ops.cast(training_value, dtypes.bool)
else:
training_value = bool(training_value)
kwargs['training'] = training_value
training_arg_passed_by_framework = True
# Only create Keras history if at least one tensor originates from a
# `keras.Input`. Otherwise this Layer may be being used outside the Keras
# framework.
if build_graph and base_layer_utils.needs_keras_history(inputs):
base_layer_utils.create_keras_history(inputs)
# Clear eager losses on top level model call.
# We are clearing the losses only on the top level model call and not on
# every layer/model call because layer/model may be reused.
if (base_layer_utils.is_in_eager_or_tf_function() and
not call_context.in_call):
self._clear_losses()
with call_context.enter(self, inputs, build_graph, training_value):
# Check input assumptions set after layer building, e.g. input shape.
if build_graph:
# Symbolic execution on symbolic tensors. We will attempt to build
# the corresponding TF subgraph inside `backend.get_graph()`
# TODO(reedwm): We should assert input compatibility after the inputs
# are casted, not before.
input_spec.assert_input_compatibility(self.input_spec, inputs,
self.name)
if (any(isinstance(x, ragged_tensor.RaggedTensor) for x in input_list)
and self._supports_ragged_inputs is False): # pylint: disable=g-bool-id-comparison
raise ValueError('Layer %s does not support RaggedTensors as input. '
'Inputs received: %s. You can try converting your '
'input to an uniform tensor.' % (self.name, inputs))
graph = backend.get_graph()
with graph.as_default(), backend.name_scope(self._name_scope()):
# Build layer if applicable (if the `build` method has been
# overridden).
self._maybe_build(inputs)
cast_inputs = self._maybe_cast_inputs(inputs)
# Wrapping `call` function in autograph to allow for dynamic control
# flow and control dependencies in call. We are limiting this to
# subclassed layers as autograph is strictly needed only for
# subclassed layers and models.
# tf_convert will respect the value of autograph setting in the
# enclosing tf.function, if any.
if (base_layer_utils.is_subclassed(self) and
not base_layer_utils.from_saved_model(self)):
call_fn = autograph.tf_convert(
self._inference, ag_ctx.control_status_ctx())
else:
call_fn = self._inference
if not self.dynamic:
try:
with base_layer_utils.autocast_context_manager(
self._compute_dtype):
# Add auto_control_deps in V2 when they are not already added by
# a `tf.function`.
if (ops.executing_eagerly_outside_functions() and
not base_layer_utils.is_in_eager_or_tf_function()):
with auto_control_deps.AutomaticControlDependencies() as acd:
outputs = call_fn(cast_inputs, *args, **kwargs)
# Wrap Tensors in `outputs` in `tf.identity` to avoid
# circular dependencies.
outputs = base_layer_utils.mark_as_return(outputs, acd)
else:
outputs = call_fn(cast_inputs, *args, **kwargs)
except errors.OperatorNotAllowedInGraphError as e:
raise TypeError('You are attempting to use Python control '
'flow in a layer that was not declared to be '
'dynamic. Pass `dynamic=True` to the class '
'constructor.\nEncountered error:\n"""\n' +
str(e) + '\n"""')
else:
# We will use static shape inference to return symbolic tensors
# matching the specifications of the layer outputs.
# Since `self.dynamic` is True, we will never attempt to
# run the underlying TF graph (which is disconnected).
# TODO(fchollet): consider py_func as an alternative, which
# would enable us to run the underlying graph if needed.
outputs = self._symbolic_call(inputs)
if outputs is None:
raise ValueError('A layer\'s `call` method should return a '
'Tensor or a list of Tensors, not None '
'(layer: ' + self.name + ').')
if base_layer_utils.have_all_keras_metadata(inputs):
if training_arg_passed_by_framework:
kwargs.pop('training')
if mask_arg_passed_by_framework:
kwargs.pop('mask')
inputs, outputs = self._set_connectivity_metadata_(
inputs, outputs, args, kwargs)
self._handle_activity_regularization(inputs, outputs)
self._set_mask_metadata(inputs, outputs, input_masks)
if hasattr(self, '_set_inputs') and not self.inputs:
# Subclassed network: explicitly set metadata normally set by
# a call to self._set_inputs().
# TODO(b/120997007): This should be done in Eager as well, but
# causes garbage collection issues because of the placeholders
# created on the default Keras graph.
self._set_inputs(inputs, outputs)
else:
# Eager execution on data tensors.
with backend.name_scope(self._name_scope()):
self._maybe_build(inputs)
cast_inputs = self._maybe_cast_inputs(inputs)
with base_layer_utils.autocast_context_manager(
self._compute_dtype):
outputs = self._inference(cast_inputs, *args, **kwargs)
self._handle_activity_regularization(inputs, outputs)
self._set_mask_metadata(inputs, outputs, input_masks)
return outputs